P5: New-release detection + back-to-back N runs to seed a fresh baseline

[FumingPower3925]

Part of #151.

Goal. When a new celeris release lands, the next weekly benchmark runs the suite 3–5× back-to-back to establish a solid baseline immediately (instead of waiting weeks for per-week samples).

Detection (the poll workflow was deleted, so no cron poller).

• Read the current celeris pin from go.mod via celerisVersion() (mage_helpers.go:196).
• Compare against a persisted 'last-benchmarked version' marker (the results/ dir is ephemeral in CI). Recommend: derive the newest benchmarked version from docs index.json (single source of truth, no extra state). Alternatives: orphan state branch (old poller pattern) or a GH variable.
• If go.mod pin != newest benchmarked version → 'new release since last benchmark'.

Back-to-back N.

• A detect job outputs is_new_release; the bench job sets BENCH_RUNS (or repeats the matrix) to N (3–5) when true, else the normal weekly N.
• Store each run under run-1..run-N in the P4: Replace the broken publish pipeline (git-committed files + tiny pointer dispatch) #165 layout; dashboard shows variance across them.
• Budget: pair the multiplier with the curated/headline matrix so N runs fit ≤24h.

Acceptance. Releasing celeris vX → the next weekly bench runs N× and publishes N runs under that version's date; subsequent weeks run 1×.

Refs: mage_helpers.go:196, deleted poll-celeris-release.yml (git history), docs index.json, mage_bench.go (BENCH_RUNS).